Pith. sign in

REVIEW 3 major objections 5 minor 35 references

Self-Organizing Edge Computing Distribution Framework for Visual SLAM

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims a three-layer self-organizing framework can fully distribute monocular ORB SLAM3's tracking, local mapping, and loop closing across networked devices and fall back to standalone single-device operation, with accuracy and…

desk verdict Genuinely useful SLAM distribution wrapper; the accuracy claim rests on runs that didn't hit unexplained state artifacts. read the letter →

arxiv 2501.08629 v1 pith:CBAGO5O4 submitted 2025-01-15 cs.RO cs.CVcs.DC

Jussi Kalliola , Lauri Suomela , Sergio Moreschini , David Hästbacka This is my paper · ORCID
classification cs.ROcs.CVcs.DC
keywords edgecomputingvisualSLAMORBSLAM3distributedsystemsself-organizingnetworksstatemanagementcollaborativekeyframe-based
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the heavy computation of visual SLAM can be spread over a group of networked devices with no central server, and that the same software can run alone on one device when the network disappears. The authors wrap monocular ORB SLAM3 in a three-layer architecture that splits its three modules—tracking, local mapping, and loop closing—into separate nodes that discover each other and agree on a distribution policy. If the claim is correct, robots with limited onboard compute could pool processing with nearby devices while remaining operational when connectivity fails. The experimental case uses public benchmark sequences and live office runs, reporting trajectory accuracy and per-node CPU use comparable to the monolithic system, with added network traffic as the cost. The paper also reports occasional state-management artifacts that inflate error in some runs, which it does not fully explain.

What carries the argument

The load-bearing mechanism is the distribution layer's heuristic policy, which decides for each node whether to offload local mapping and loop closing to discovered nodes or run them locally, together with a two-tier state model that keeps an eventually consistent copy of the map across nodes. A three-layer architecture—core SLAM, distribution, and communication—keeps changes to the SLAM engine minimal, and the communication layer lets keyframe and map updates double as the offloading signal. This combination of policy and state management is what lets the system self-organize without a server and degrade to standalone operation.

What would settle it

Run the one-device configuration and unmodified ORB SLAM3 on the Machine Hall 04 and Room 1 benchmark sequences with identical keyframe parameters; if trajectory error stays near the original 0.080 m and 0.090 m, the core modifications are exonerated, and if it does not, the clean fallback claim fails. Then rerun the three-device configuration on the Machine Hall 03 sequence ten times with an explicit pause-acknowledgement before global map updates and check whether the artifact-driven error of 0.818 m reappears.

Watch

Extended reading notes

Core claim

The central claim is that the three parallel threads of a keyframe-based SLAM system can be treated as distributable units and orchestrated by a simple discovery-based policy, producing a system that executes the full pipeline cooperatively on several devices or alone on one. Implemented as a minimally invasive wrapper around monocular ORB SLAM3, the framework routes tracking output to whichever node is discovered to host local mapping and loop closing, and falls back to executing all modules locally when no partner is found. The paper argues that the distributed version matches the monolithic baseline's accuracy and resource utilization in most benchmark runs, with CPU usage slightly lower and bandwidth consumption added, and that single-device mode performs like standalone ORB SLAM3 with a few observable discrepancies. The paper's own data shows those discrepancies: some single-device sequences have markedly higher trajectory error than the original, and some multi-device runs produce unexplained trajectory artifacts after global map updates.

Load-bearing premise

Everything rests on the assumption that the paper's deliberate changes to ORB SLAM3's keyframe creation and startup behavior do not meaningfully change tracking quality; when the whole system runs on one device, several sequences end up with much higher trajectory error than the original system, so that assumption is not clearly satisfied.

Editorial extensions

If this is right

  • If the claim holds, a team of heterogeneous robots can run a full SLAM pipeline collaboratively without any central server or fixed role assignment.
  • The same framework can fall back to standalone operation on a single device, so losing all network peers does not halt localization.
  • Offloading local mapping and loop closing to more powerful edge nodes lowers per-device CPU utilization by roughly 1 to 6 percent in the reported experiments, at the cost of sustained network bandwidth.
  • Because the core is wrapped rather than rewritten, the distribution design could be adapted to other modular SLAM systems with the same three-thread structure.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the distribution policy is only as responsive as node discovery; if discovery is slow or a node vanishes mid-run, the fallback branches may not engage in time to prevent tracking failure.
  • A direct test of the core modifications' impact would be to run unmodified ORB SLAM3 with the same keyframe threshold and spacing; if the trajectory error gaps on single-device runs persist, they come from the wrapper, not the parameter changes.
  • The nondeterministic artifact pattern suggests a testable fix: require an explicit acknowledgement from every node before a global map update takes effect, then measure whether artifact frequency drops.
  • The wrapper design could plausibly extend to stereo or visual-inertial SLAM systems whose back-end threads are similarly separable, but that extension is not demonstrated in the paper.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a self-organizing edge-computing framework for visual SLAM that can distribute the three ORB SLAM3 modules (tracking, local mapping, loop closing) across heterogeneous networked devices, or run all modules on a single device when no network is available. The framework is implemented as a three-layer wrapper around monocular ORB SLAM3, with a distribution policy, an observer-based state management model, and ROS2/FastDDS communication. The authors evaluate the system on EuRoC, TUM, and real-world office datasets in both 1-node and 3-node configurations, comparing ATE, CPU utilization, failure rate, network bandwidth, and message frequency against monolithic ORB SLAM3. The central claim is that the distributed system matches the accuracy and resource utilization of the monolithic approach while enabling collaborative execution and graceful degradation.

Significance. If the claims are substantiated, the framework is a useful engineering contribution: it addresses a real limitation of client-server edge SLAM systems by providing a device-agnostic, failure-resilient distribution layer, and it is demonstrated on a widely used SLAM system. The paper includes real-world experiments, network analysis, and a discussion of state-management issues, which is commendable. However, the evaluation as presented does not fully establish the central claims because the accuracy comparison relies on selectively excluding runs with unexplained, nondeterministic artifacts, and because the core SLAM system was modified in ways that affect the baseline comparison. The resource-utilization claim is also based on a partial metric. The framework itself and the experimental design are valuable, but the evidence needs to be substantially strengthened or the claims appropriately qualified.

major comments (3)
  1. [Section VI-A, Tables Ia/Ib] The claim that the proposed framework 'matches the accuracy ... of the monolithic approach' is supported only by the bold table entries, which the caption defines as experiments where no sequence had max ATE ≥ 1.0 m due to state-management issues. The text reports that these artifacts are nondeterministic and could not be reproduced (Section VI-A) and that their investigation is left to future work (Section VI-C). Because the tabulated averages mix clean and artifact-affected runs (e.g., MH03 3-node ATE 0.176 is the average of 0.040, 0.818, 0.037), the central accuracy claim is conditional on a stochastic event that is not characterized. The authors should either eliminate these artifacts, provide a reproducible analysis of their causes, or substantially qualify the claim to reflect the observed run-to-run variance.
  2. [Section IV-A and Table I (1-node results)] The core ORB SLAM3 system was modified in ways that directly affect tracking quality: the keyframe creation threshold was changed from 90% to 80%, a minimum of two frames between keyframes was imposed, and the initialization wait behavior was altered. The 1-node configuration is intended to demonstrate clean degradation to standalone SLAM, but the reported ATE values differ substantially from ORB SLAM3 on several sequences (e.g., EuRoC MH04: 0.533 m vs 0.080 m; V2 02: 0.191 m vs 0.084 m; TUM Room1: 0.239 m vs 0.090 m). These differences show that the comparison against the monolithic baseline is confounded by the core modifications. The authors should either justify that these modifications are accuracy-neutral (e.g., by comparing against ORB SLAM3 with the same modifications) or explicitly acknowledge that the 1-node results do not represent unmodified ORB SLAM3 behavior.
  3. [Table I caption and Section V-D] The resource-utilization comparison is incomplete and potentially misleading. The table caption states that 'all metrics are measured at TR node,' so the reported CPU utilization reflects only the tracking node's process. In the 3-node configuration, the offloaded LM and LC modules consume CPU on other machines; the total system-wide CPU usage across all nodes is not reported. Therefore, the claim that the distributed system 'matches ... resource utilization of the monolithic approach' is not established from the data presented. The authors should report CPU utilization for the LM and LC nodes (and a total) to support the resource-utilization claim, or change the claim to refer only to tracking-node CPU.
minor comments (5)
  1. [Section VI-A, paragraph 1] The sentence 'Even in cases of failed state management, we consider that the distribution system can execute SLAM in a fully distributed setting with comparable performance to the original system since the failure cases are highly irregular' appears to assert the conclusion the experiments are meant to demonstrate. Please rephrase to present this as an interpretation or hypothesis, not as a justification.
  2. [Section V-D] The metric 'Fails' is not defined. It would be helpful to state exactly what constitutes a tracking failure and how the rate is computed (e.g., fraction of frames lost, number of lost-track events, or something else).
  3. [Section III-C] The definitions of the state equations contain typographical issues, for example the expression for S^i_full is garbled. Please rewrite the formal definitions of S_full and S_slam more clearly.
  4. [Table Ib] Several real-life experiment entries are marked with an asterisk and the meaning is given in a footnote, but the table body also contains entries without a numeric value. Please ensure all missing or invalid entries are either populated or explicitly annotated.
  5. [References [32], [33]] The TUM datasets used for the 'TUM Office' experiments should be cited precisely. Reference [33] is the TUM-VI visual-inertial benchmark, which may not be the same as the monocular office sequences used here; please clarify the exact dataset and sequence names.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is an empirical comparison against external benchmarks and an independently published baseline, not a derivation from fitted inputs or self-citations.

full rationale

The paper's central claim, that the proposed DSLAM framework matches ORB SLAM3 accuracy and resource utilization, is supported by benchmark experiments on EuRoC, TUM, and real-world office trajectories, with ATE measured against ground truth or the original ORB SLAM3 trajectory. No parameter is fitted to the target metric, and no 'prediction' is produced from the same data used to evaluate it. The distribution policy in Section III-D is a simple heuristic over discovered nodes, not a fitted model. The core-layer modifications in Section IV-A change keyframe thresholds and initialization behavior, but these are disclosed algorithmic changes rather than inputs that force the reported outcomes by construction. The paper explicitly acknowledges limitations in Section VI-A and VI-C, including artifacts whose cause it 'could not identify' and which it 'was unable to reproduce in a deterministic manner,' and it leaves investigation of map initialization and global map update failures to future work. These are significant threats to the strength and generalizability of the empirical claim, and the selective emphasis on bold table entries is a reporting concern, but none of these issues makes the derivation circular. There are no load-bearing self-citations by the authors. The evaluation is self-contained against external benchmarks, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on three hand-tuned communication and keyframe parameters, plus the domain assumption that a three-thread SLAM core can be distributed with eventual consistency. No new physical or algorithmic entities are postulated. The main uncharged assumptions are that the core modifications are behaviorally neutral and that eventual consistency is sufficient, both of which are challenged by the paper's own results.

free parameters (3)
  • Keyframe creation threshold = reference map point ratio 80% (from 90%); at least 2 frames between keyframes
    Set in Section IV-A to reduce keyframe rate. This tuning changes the behavior of the core SLAM system and directly affects the accuracy comparison against unmodified ORB SLAM3.
  • Local map batch size and delay = 3 to 15 keyframes per batch, 50 ms between updates
    Set in Section V-A to control network congestion and state propagation. These values are hand-chosen and no sensitivity analysis is provided.
  • Global map batch size and delay = 10 keyframes per batch, 100 ms between updates
    Set in Section V-A for global map updates after GBA, loop closing, or map merge. The choice affects the probability of stale-state artifacts during global updates.
assumptions (4)
  • domain assumption The core SLAM system is divided into three parallel threads dedicated to short-term (tracking), mid-term (local mapping), and long-term (loop closing) data associations.
    Invoked in Section III as the precondition for the whole distribution scheme. It holds for ORB SLAM3, but limits the claimed device-agnostic generality to SLAM systems with this thread decomposition.
  • domain assumption Eventual consistency is sufficient for correct distributed SLAM operation; nodes may run in soft-state until consistency is achieved.
    Adopted in Section III-C from distributed systems theory. The paper assumes that eventual consistency will not corrupt the SLAM map in practice, yet the observed artifacts show this assumption is not always satisfied.
  • domain assumption The only behavioral differences between the distributed system's core and the original ORB SLAM3 are the modifications listed in Section IV-A.
    Needed for the 1-node experiments to be interpreted as a clean degradation to standalone ORB SLAM3. The large 1-node ATE differences in several datasets suggest this assumption is questionable.
  • domain assumption ROS2 reliable QoS and the described connectivity topology deliver state updates within latency bounds that preserve real-time tracking.
    Used throughout Section V-A and the connectivity scheme in Section III-B. The paper does not measure latency directly, only aggregate bandwidth and message frequency.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Organizing Edge Computing Distribution Framework for Visual SLAM." pith.science (2026). https://pith.science/paper/CBAGO5O4

@misc{pith2026250108629,
  author       = {Pith},
  title        = {Pith review of: Self-Organizing Edge Computing Distribution Framework for Visual SLAM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CBAGO5O4}},
  note         = {Machine review of arXiv:2501.08629}
}
read the original abstract

Localization within a known environment is a crucial capability for mobile robots. Simultaneous Localization and Mapping (SLAM) is a prominent solution to this problem. SLAM is a framework that consists of a diverse set of computational tasks ranging from real-time tracking to computation-intensive map optimization. This combination can present a challenge for resource-limited mobile robots. Previously, edge-assisted SLAM methods have demonstrated promising real-time execution capabilities by offloading heavy computations while performing real-time tracking onboard. However, the common approach of utilizing a client-server architecture for offloading is sensitive to server and network failures. In this article, we propose a novel edge-assisted SLAM framework capable of self-organizing fully distributed SLAM execution across a network of devices or functioning on a single device without connectivity. The architecture consists of three layers and is designed to be device-agnostic, resilient to network failures, and minimally invasive to the core SLAM system. We have implemented and demonstrated the framework for monocular ORB SLAM3 and evaluated it in both fully distributed and standalone SLAM configurations against the ORB SLAM3. The experiment results demonstrate that the proposed design matches the accuracy and resource utilization of the monolithic approach while enabling collaborative execution.

Figures

Figures reproduced from arXiv: 2501.08629 by the authors.

Figure 1
Figure 1. High-level illustration of the proposed distribution [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Connectivity scheme. Arrows indicate the direction [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Distribution system architecture. The tracking module was modified to reduce the rate of keyframe construction. Keyframes are now created only if at least two frames have elapsed since the last keyframe, and the reference map point ratio was adjusted from 90% to 80%, requiring a higher number of tracked points for keyframe generation. Additionally, initialization process was adjusted to wait until the initial keyfra… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: State managers for Keyframes and Maps. A dashed outline indicates that data is transmitted to another module/layer. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Artifacts can be seen between timesteps 200-220. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 30 canonical work pages

  1. [1]

    COVINS: visual-inertial SLAM for centralized collaboration,

    P. Schmuck, T. Ziegler, M. Karrer, J. Perraudin, and M. Chli, “COVINS: visual-inertial SLAM for centralized collaboration,” CoRR, vol. abs/2108.05756, 2021

    arXiv 2021

  2. [2]

    Ccm-slam: Robust and efficient centralized collaborative monocular simultaneous localization and mapping for robotic teams,

    P. Schmuck and M. Chli, “Ccm-slam: Robust and efficient centralized collaborative monocular simultaneous localization and mapping for robotic teams,” Journal of Field Robotics, vol. 36, no. 4, pp. 763–781, 2019

    work page 2019

  3. [3]

    Distributed and collaborative monocular simultaneous localization and mapping for multi-robot sys- tems in large-scale environments,

    H. Zhang, X. Chen, H. Lu, and J. Xiao, “Distributed and collaborative monocular simultaneous localization and mapping for multi-robot sys- tems in large-scale environments,” International Journal of Advanced Robotic Systems, vol. 15, no. 3, p. 1729881418780178, 2018

    work page 2018

  4. [4]

    Burgard, O

    W. Burgard, O. Brock, and C. Stachniss, Mapping Large Loops with a Single Hand-Held Camera , 2008, pp. 297–304

    work page 2008

  5. [5]

    Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. M. Montiel, and J. D. Tard ´os, “Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,” IEEE Transactions on Robotics , vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  6. [6]

    Vins-mono: A robust and versatile monoc- ular visual-inertial state estimator,

    T. Qin, P. Li, and S. Shen, “Vins-mono: A robust and versatile monoc- ular visual-inertial state estimator,” IEEE Transactions on Robotics , vol. 34, no. 4, pp. 1004–1020, 2018

    2018

  7. [7]

    Lift-slam: A deep-learning feature-based monocular visual slam method,

    H. M. S. Bruno and E. L. Colombini, “Lift-slam: A deep-learning feature-based monocular visual slam method,” Neurocomputing, vol. 455, pp. 97–110, 2021

    work page 2021

  8. [8]

    Edge-slam: Edge-assisted visual simultaneous localization and mapping,

    A. J. Ben Ali, M. Kouroshli, S. Semenova, Z. S. Hashemifar, S. Y . Ko, and K. Dantu, “Edge-slam: Edge-assisted visual simultaneous localization and mapping,” ACM Trans. Embed. Comput. Syst., vol. 22, no. 1, oct 2022

    work page 2022

Show all 35 references
  1. [9]

    {SwarmMap}: Scaling up real-time collaborative visual {SLAM} at the edge,

    J. Xu, H. Cao, Z. Yang, L. Shangguan, J. Zhang, X. He, and Y . Liu, “{SwarmMap}: Scaling up real-time collaborative visual {SLAM} at the edge,” in 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22) , 2022, pp. 977–993

    2022

  2. [10]

    edgeslam2: Rethinking edge-assisted visual slam with on-chip intel- ligence,

    D. Li, Y . Zhao, J. Xu, S. Zhang, L. Shangguan, and Z. Yang, “edgeslam2: Rethinking edge-assisted visual slam with on-chip intel- ligence,” in IEEE INFOCOM 2024 - IEEE Conference on Computer Communications, 2024, pp. 1481–1490

    2024

  3. [11]

    Edge assisted mobile semantic visual slam,

    H. Cao, J. Xu, D. Li, L. Shangguan, Y . Liu, and Z. Yang, “Edge assisted mobile semantic visual slam,” IEEE Transactions on Mobile Computing, vol. 22, no. 12, pp. 6985–6999, 2023

    2023

  4. [12]

    Dynnetslam: Dynamic visual slam network offloading,

    P. Sossalla, J. Hofer, J. Rischke, C. Vielhaus, G. T. Nguyen, M. Reisslein, and F. H. P. Fitzek, “Dynnetslam: Dynamic visual slam network offloading,” IEEE Access, vol. 10, pp. 116 014–116 030, 2022

    2022

  5. [13]

    Cloud- slam: Edge offloading of stateful vehicular applications,

    K.-L. Wright, A. Sivakumar, P. Steenkiste, B. Yu, and F. Bai, “Cloud- slam: Edge offloading of stateful vehicular applications,” in 2020 IEEE/ACM Symposium on Edge Computing (SEC) , 2020, pp. 139– 151

    2020

  6. [14]

    Of- floading visual slam processing to the edge: An energy perspective,

    P. Sossalla, J. Hofer, C. Vielhaus, J. Rischke, and F. H. P. Fitzek, “Of- floading visual slam processing to the edge: An energy perspective,” in 2023 International Conference on Information Networking (ICOIN) , 2023, pp. 39–44

    2023

  7. [15]

    Estimating uncertain spatial relationships in robotics,

    R. Smith, M. Self, and P. Cheeseman, “Estimating uncertain spatial relationships in robotics,” in Uncertainty in Artificial Intelligence , ser. Machine Intelligence and Pattern Recognition, J. F. LEMMER and L. N. KANAL, Eds. North-Holland, 1988, vol. 5, pp. 435–461

    1988

  8. [16]

    On the representation and estimation of spatial uncertainty,

    R. C. Smith and P. Cheeseman, “On the representation and estimation of spatial uncertainty,”The International Journal of Robotics Research, vol. 5, no. 4, pp. 56–68, 1986

    1986

  9. [17]

    Covins-g: A generic back-end for collaborative visual-inertial slam,

    M. Patel, M. Karrer, P. B ¨anninger, and M. Chli, “Covins-g: A generic back-end for collaborative visual-inertial slam,” 2023

    2023

  10. [18]

    Real-time simultaneous localisation and mapping with a single camera,

    Davison, “Real-time simultaneous localisation and mapping with a single camera,” in Proceedings Ninth IEEE International Conference on Computer Vision , 2003, pp. 1403–1410 vol.2

    2003

  11. [19]

    Monoslam: Real-time single camera slam,

    A. J. Davison, I. D. Reid, N. D. Molton, and O. Stasse, “Monoslam: Real-time single camera slam,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 29, no. 6, pp. 1052–1067, 2007

    2007

  12. [20]

    Inverse depth parametrization for monocular slam,

    J. Civera, A. J. Davison, and J. M. M. Montiel, “Inverse depth parametrization for monocular slam,” IEEE Transactions on Robotics, vol. 24, no. 5, pp. 932–945, 2008

    2008

  13. [21]

    Orb-slam: A versatile and accurate monocular slam system,

    R. Mur-Artal, J. M. M. Montiel, and J. D. Tard ´os, “Orb-slam: A versatile and accurate monocular slam system,” IEEE Transactions on Robotics, vol. 31, no. 5, pp. 1147–1163, 2015

    2015

  14. [22]

    Orb-slam2: an open-source slam system for monocular, stereo and rgb-d cameras,

    R. Mur-Artal and J. Tardos, “Orb-slam2: an open-source slam system for monocular, stereo and rgb-d cameras,” IEEE Transactions on Robotics, vol. PP, 10 2016

    2016

  15. [23]

    Cudasift-slam: multiple-map visual slam for full procedure mapping in real human endoscopy,

    R. Elvira, J. Tard ´os, and J. Montiel, “Cudasift-slam: multiple-map visual slam for full procedure mapping in real human endoscopy,” 05 2024

    2024

  16. [24]

    1-point ransac for extended kalman filtering: Application to real-time structure from motion and visual odometry,

    J. Civera, O. Grasa, A. Davison, and J. Montiel, “1-point ransac for extended kalman filtering: Application to real-time structure from motion and visual odometry,” J. Field Robotics, vol. 27, pp. 609–631, 09 2010

    2010

  17. [25]

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,

    Z. Teed and J. Deng, “Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,” Advances in neural information processing systems, vol. 34, pp. 16 558–16 569, 2021

    2021

  18. [26]

    Bags of binary words for fast place recognition in image sequences,

    D. Galvez-L ´opez and J. D. Tardos, “Bags of binary words for fast place recognition in image sequences,” IEEE Transactions on Robotics , vol. 28, no. 5, pp. 1188–1197, 2012

    2012

  19. [27]

    Parallel tracking and mapping for small ar workspaces,

    G. Klein and D. Murray, “Parallel tracking and mapping for small ar workspaces,” in 2007 6th IEEE and ACM International Symposium on Mixed and Augmented Reality , 2007, pp. 225–234

    2007

  20. [28]

    Distributed mapping with privacy and communication constraints: Lightweight algorithms and object-based models,

    S. Choudhary, L. Carlone, C. Nieto, J. Rogers, H. I. Christensen, and F. Dellaert, “Distributed mapping with privacy and communication constraints: Lightweight algorithms and object-based models,” The International Journal of Robotics Research, vol. 36, no. 12, pp. 1286– 1311, 2017

    2017

  21. [29]

    AdaptSLAM: Edge-assisted adaptive slam with resource constraints via uncertainty minimization,

    Y . Chen, H. Inaltekin, and M. Gorlatova, “AdaptSLAM: Edge-assisted adaptive slam with resource constraints via uncertainty minimization,” in Proc. IEEE INFOCOM , 2023

    2023

  22. [30]

    M. v. Steen and A. S. Tanenbaum, Distributed Systems, fourth edition, version 4.01 (january 2023) ed. Erscheinungsort nicht ermittelbar: Maarten van Steen, 2023

    2023

  23. [31]

    The euroc micro aerial vehicle datasets,

    M. Burri, J. Nikolic, P. Gohl, T. Schneider, J. Rehder, S. Omari, M. W. Achtelik, and R. Siegwart, “The euroc micro aerial vehicle datasets,” The International Journal of Robotics Research , 2016

    2016

  24. [32]

    Tum-vie: The tum stereo visual-inertial event dataset,

    S. Klenk, J. Chui, N. Demmel, and D. Cremers, “Tum-vie: The tum stereo visual-inertial event dataset,” in International Conference on Intelligent Robots and Systems (IROS) , 2021

    2021

  25. [33]

    The tum vi benchmark for evaluating visual-inertial odometry,

    D. Schubert, T. Goll, N. Demmel, V . Usenko, J. Stueckler, and D. Cremers, “The tum vi benchmark for evaluating visual-inertial odometry,” in International Conference on Intelligent Robots and Systems (IROS), October 2018

    2018

  26. [34]

    A benchmark for the evaluation of rgb-d slam systems,

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, “A benchmark for the evaluation of rgb-d slam systems,” in 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2012, pp. 573–580

    2012

  27. [35]

    evo: Python package for the evaluation of odometry and slam

    M. Grupp, “evo: Python package for the evaluation of odometry and slam.” https://github.com/MichaelGrupp/evo, 2017

    2017

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.